The Wind Still Howls Where No One Stands | Monolith Monthly
Every step is worth recording.
Monolith Monthly is our recurring column where we track Monolith's progress alongside that of our portfolio companies — preserving a clear sense of process along a finite timeline, and marking the moments and developments worth watching.
01｜Portfolio Updates
Key developments from our portfolio companies: funding rounds, IPOs, strategic partnerships, technical breakthroughs, and product launches — the moments that matter.
02｜Monolith Moment
Spotlighting Monolith's milestones, honors, and perspectives — recording the convictions and actions we stand by.
01 Portfolio Updates
Moonshot AI Releases Open-Source Audio Model Kimi-Audio
Moonshot AI recently released Kimi-Audio, a general-purpose audio foundation model supporting automatic speech recognition (ASR), audio question answering (AQA), automatic audio captioning (AAC), speech emotion recognition (SER), sound event/scene classification (SEC/ASC), and end-to-end speech dialogue. Kimi-Audio uses hybrid audio inputs (continuous acoustic vectors + discrete semantic tokens) with an LLM backbone featuring parallel heads to generate text and audio tokens, and employs a flow-matching-based chunked streaming detokenizer for low-latency audio generation. It achieved SOTA results across multiple audio benchmarks.
Muyan Intelligence Releases Open-Source Model Muyan-TTS
Muyan Intelligence recently released Muyan-TTS, an open-source text-to-speech model. The model is cost-efficient with strong support for secondary development, delivers excellent English performance, and achieves industry-leading inference speed — generating one second of speech in just 0.33 seconds, the fastest among all open-source models. Its word error rate and speech quality scores rank among the top tier of open-source models. Muyan Intelligence released two variants: a base model pre-trained on diverse long-form audio datasets for zero-shot TTS synthesis, and a supervised fine-tuning (SFT) model trained on a single speaker to enhance TTS performance. Detailed training methods and a framework for optimizing inference efficiency are also provided.
StarMatrix Innovation's Products Win Red Dot Design Award
StarMatrix Innovation's Beatbot brand announced that its next-generation flagship products, AquaSense 2 Ultra and AquaSense 2, have won the 2025 Red Dot Design Award. The Red Dot Design Award is one of the world's most prestigious design honors, recognizing works of exceptional innovation and quality. The AquaSense 2 Ultra integrates biomimetic inspiration with fluid dynamics aesthetics, featuring a sculptural unibody design and HybridSense™ AI Pool Mapping that adapts to every pool detail. The AquaSense 2 uses IMR工艺 with a 4H scratch-resistant coating, achieving an ideal balance between durability and refined texture.
Lxson's WAWA Robot Debuts at JD.com Event
Lxson recently announced the completion of its angel and angel+ rounds totaling nearly RMB 300 million. The company's first prototype, "WAWA," featuring emotional interaction capabilities, made its debut at JD.com's 618 Summer Concert event, where it was auctioned starting at RMB 1 and drew nearly 40,000 online viewers. The company stated that future iterations of WAWA will be fully equipped with a consciousness-biomimetic model, with customizable voice and personality features gradually rolled out to strengthen its household interaction attributes.
MetaX Partners with HuaKe ZhiGu on New AI Large Model Appliance
MetaX Integrated Circuit has partnered with HuaKe ZhiGu to launch a liquid-cooled AI large model appliance. The product features two-phase self-driven loop heat pipe (LHP) liquid cooling technology, offering maintenance-free operation, zero leakage risk, ultra-quiet performance, and precise temperature control. It is equipped with MetaX's XiCloud C-series high-performance general-purpose GPUs, supporting mixed-precision computing and inter-card interconnect technology to handle high-load scenarios including AI model fine-tuning, large model inference, and generative AI. The device comes pre-installed with multiple models including the full-parameter DeepSeek-R1, DeepSeek-70B, and DeepSeek-32B.
D-Robotics Enables Efficient Motion Control for Unitree's G1 Humanoid Robot
Building on optimizations from Horizon Robotics' lab paper Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills, D-Robotics successfully deployed its optimized ASAP framework on the RDK S100. The framework employs a two-stage training strategy combined with a Delta action model to achieve efficient whole-body motion control for humanoid robots in real-world environments, with stable and outstanding physical performance. This breakthrough significantly reduces deployment resource overhead: model inference occupies only 2% of BPU capacity, while CPU usage drops by 250% compared to pure CPU inference — freeing up more compute for more complex tasks such as visual detection, object recognition, path navigation, and intelligent decision-making.
GravityXR Unveils EB100 Chip
At the 15th Songshan Lake China IC Innovation Summit, GravityXR introduced the EB100, a low-power spatial rendering and display chip designed for robotics/XR applications. The chip can generate rendered facial robot avatars and enable real-time AI-driven interaction, allowing robots to clearly express facial emotions and transmit the operator's emotions in real time. Key differentiated features include real-time 3D model reconstruction and driving, 3D display image processing and enhancement, multi-view platform rendering, multi-layer compositing, and interfaces including PCIe, USB 3.0, and MIPI.
Infinigence AI Launches DeepSeek-R1-0528 Inference Acceleration Version
Infinigence AI recently announced the launch of the DeepSeek-R1-0528 inference acceleration version. The new version builds on the DeepSeek V3 Base released last December, with increased compute investment during post-training that significantly enhances the model's reasoning depth and capability. The model has 685B parameters (including 14B MTP layers). Official data shows the new model reduces hallucination rates by approximately 45-50%, delivering more accurate and reliable results, with improved capabilities in front-end code generation, role-playing, and other domains.
02 Monolith Moment
Monolith Ranked Among Top 10 Most Active Investors in Embodied Intelligence
Diyixinsheng recently released its Q1 2025 China Most Active Investment Institutions series rankings, with Monolith selected for:
Top 30 Most Active VC Firms in China,
Top 10 Most Active Investors in Embodied Intelligence.
Monolith Named to Top 20 Most Popular Direct Investment Funds Among Young Investors
The China Venture Capital and Private Equity Fund of Funds Research Center released its 2025 Most Popular Direct Investment Funds ranking,
with Monolith selected for:
2025 Top 20 Most Popular Direct Investment Fund Firms Among Young Investors;
Monolith's Xi Cao Debuts on Forbes 2025 Midas List at No. 39
The Forbes Midas List, often called the "Oscars" of the investment world, recently unveiled its 2025 rankings. Xi Cao, founding partner of Monolith, made his debut on the list at No. 39 globally, recognized for his early investments in Kuaishou, Unitree, and other companies.
Cao was also named to the Rongzhong U45 China's Best Young Investors and the Fund of Funds Research Center's "40U40" Best Direct Investment Fund Young Investors lists.
